SERS Mixture Recognition from Pure-Substance Spectra via Component Evidence Learning and Two-Stage Inference

Surface-enhanced Raman spectroscopy (SERS) is a powerful tool for molecular analysis, yet the recognition of mixed spectra remains challenging because severe peak overlap makes mixture-specific data expensive to acquire and difficult to cover exhaustively. Current machine-learning approaches often rely on labeled mixture datasets, synthetic mixed spectra, or prior component-matching schemes, making their performance strongly dependent on task-specific mixture data. A pure-spectrum-trained framework for SERS mixture recognition is presented here based on component evidence learning and two-stage inference. Using paraquat, thiram, and tricyclazole as representative target compounds, the framework learns reusable constituent-level evidence directly from pure-substance spectra and converts it into mixture-category predictions within a unified recognition model. This design avoids mixture-specific parameter training while enabling direct recognition of binary and ternary mixtures. Experiments on SERS spectral datasets yielded a mixture recognition accuracy of 98.58%. The results show that pure-substance spectral learning can support accurate recognition of complex SERS mixtures and provide a scalable strategy for mixture analysis when labeled mixture data are limited.